Multi-Agent LLM Systems for Trading

Large language models (LLMs) are evolving rapidly, from early text-generation systems to the latest models capable of sophisticated reasoning and multi-step decision-making. Reference [1] applies LLMs to the development of a gold trading system. Unlike traditional quantitative models, which are typically built around forecasting future prices or returns, the authors use LLMs as a committee of analysts rather than as predictive models.

The objective is to test whether structured LLM reasoning can generate better trading decisions than conventional indicator-based approaches. Specifically, the framework consists of three agents responsible for data analysis, risk assessment, and trade decision-making. The authors employ Chain-of-Thought prompting and evaluate model performance across temperature settings ranging from 0.0 to 1.0. They pointed out,

In this study, we analyzed whether LLM-assisted trading strategies could achieve improved performance compared to traditional indicator-based strategies when applied to historical gold market data. More specifically, we examined how LLM-assisted strategies performed compared to baseline strategies under the same market conditions.

…The results suggest that LLM-assisted strategies outperformed the baseline strategies in terms of total return. The LLM-based strategies achieved returns ranging from 33.07% to 61.17%, while the best-performing baseline strategy reached 17.53%. Additionally, all LLM-assisted strategies achieved a Sharpe ratio above 1.0, whereas none of the baseline strategies surpassed that threshold. The temperature setting t = 0.5 resulted in the best overall performance and achieved the highest values across most performance metrics. This indicates that a moderate level of randomness in the LLM output may contribute positively to trading performance.

In short, the paper finds that all LLM-based strategies outperform the traditional technical-indicator benchmarks in terms of total return, while achieving Sharpe ratios above 1.0.

Although the study has several limitations, including a single asset, only 294 trading days of data, the absence of transaction costs, no incorporation of news or sentiment inputs, and no out-of-sample tests, it nevertheless offers an interesting direction for trading-system development. Rather than serving solely as forecasting engines, LLMs may be used as autonomous agents or as specialized decision-making modules within a broader trading framework.

Let us know what you think in the comments below or in the discussion forum.

References

[1] Phillips, A., & Emilsson, V. (2026), LLM-Assisted Gold Trading: Evaluating Reasoning-Based Strategies Against Technical Indicator Baselines, Bachelor's thesis, Department of Computer and Systems Sciences, Stockholm University.

Originally Published Here: Multi-Agent LLM Systems for Trading

source https://harbourfronts.com/multi-agent-llm-systems-trading/